Summary

图 0-1｜四层市场结构：Harness 运行时层的定位

数据来源：基于本文分析绘制的示意图。

1. Overall Assessment: The Current Stage of AI Harness Development

1.1. Stage Determination

Assessment: AI Harness is currently at the stage where "the third-generation architecture has been established, the standardization window is closing, and the methodology has not yet converged."

This assessment comprises three clauses, and all three are indispensable:

1.2. Basis for the Assessment

Basis One · Leading vendors have written Harness into their official vocabulary.

In 2025 Anthropic used "harness" to describe the Claude Agent SDK; Mitchell Hashimoto named "Harness Engineering" on 2026-02-05; OpenAI published Harness engineering on 2026-02-11; and martinfowler.com published a methodology-architecture article in February 2026. Four independent sources converged within the same month, a marker that the concept has matured.

Basis Two · There is already engineering evidence at the million-line-of-code scale.

OpenAI's experiment (Grade A): roughly 1 million lines of code, about 1,500 PRs, 3 engineers (later expanded to 7), 3.5 PRs/person/day, time cost roughly 1/10 of handwritten work, 0 lines of human-written code, and single-task runs often exceeding 6 hours.

Basis Three · There are already quantifiable governance benefits.

Anthropic's internal sandbox usage: a 84% reduction in permission prompts (Grade A).

Basis Four · The ecosystem has already reached infrastructure scale.

10,000+ public MCP servers, 97 million monthly downloads of the Python + TypeScript SDKs, and AGENTS.md adopted by 60,000+ open-source projects (Grade A).

Basis Five · Standardization organizations have become involved.

AAIF (2025-12-09) and A2A donated to the Linux Foundation (2025-06-23); China established a seven-part closed loop with GB/Z 185—2026.

1.3. Four Manifestations of the Stage's Characteristics

2. Eight Core Propositions

2.1. The Bottleneck Has Already Moved Outward

Proposition: The bottleneck constraining agent-system performance has shifted from inside the model to the engineering environment outside the model.

Support:

2.2. Harness Turns Uncertainty into Predictability

Proposition: Harness's defining responsibility is to narrow the model's probabilistic output down to an interval acceptable to engineering.

Support: A unified parameter-card definition + martinfowler.com's delineation (Harness is a collection of specifications, quality checks, and workflow guidance — an intermediate layer that can be continuously improved) + Anthropic sandbox data (permission prompts reduced by 84%, while safety improved).

Quantitative Anchor: 84% (Grade A).

2.3. Verifiability Is Harness's First Value

Proposition: Harness turns "model capability" from a subjective impression into an engineering object that can be verified through regression.

Support:

2.4. Governance and Autonomy Are a Positive-Sum Game, Not a Trade-off

Proposition: Constraint is not the opposite of autonomy; it is the precondition for autonomy.

Support:

2.5. The Standardization Window Is Closing

Proposition: The standardization of agent interfaces has moved from "a hundred schools contending" into the stage of "neutral governance + locking in de facto standards."

Support:

2.6. Adoption Has Saturated, but Trust Has Not Caught Up

Proposition: The gap between adoption rate and trust is Harness's market opportunity.

Support (Grade A, Stack Overflow 2025):

Supplementary (Grade B, ): DORA 2025 shows AI adoption positively correlated with software delivery throughput and product performance, but negatively correlated with delivery stability.

2.7. Three Shifts Outward: From Prompt to Context to Harness

Proposition: The migration of focus is a layered accumulation, not a replacement of one by the other.

Not a Replacement, but Layered Accumulation: Writing prompts today is still necessary — it is simply no longer sufficient.

2.8. Harness Will Become Lighter, but It Will Not Disappear

Proposition: Stronger models will trim away part of the Harness design, but they will not eliminate Harness.

Support:

• Anthropic officially admits it is "dynamically trimming redundant Harness designs as model capabilities improve" (Grade A);
• At the same time, the post-training coupling between the model and a specific Harness has produced the situation where "changing the tool logic actually makes the model perform worse" — which reinforces the importance of Harness design (Grade A);
• martinfowler.com: "A good harness should not necessarily aim to fully eliminate human input, but to direct it to where our input is most important." (Grade A)

Conclusion: Harness will become lighter, and its composition will change, but it will not disappear.

3. Implications for the Six Industry Groups

3.1. Overall Implications

No matter which industry group, deploying AI Harness should follow the same path, with only the concrete form of each layer differing:

One cross-industry general principle: Build L5 (evaluation) first, then expand L2 (tools). Without criteria you cannot tell whether adding or removing tools is effective — the reason Vercel dared to cut its tools from 15 to 2 is precisely that they had an accuracy criterion (Grade C, ).

3.2. Software Engineering Group

• Maturity: The highest. All official evidence for the third-generation architecture (Claude Code, Codex CLI, OpenAI's million-line experiment) comes from this domain.
• Key implication: Treat AGENTS.md as a "map" rather than an "encyclopedia" (about 100 lines + a structured docs/), and write linter error messages as repair instructions injected into agents.
• What most needs strengthening: The L5 evaluation layer — moving from "passing tests" to "trajectory-level evaluation and Harness coverage."
• Risk: Over-reliance on the specific Harness structure that is post-training coupled; "bitter lesson"-style rework caused by hard-coded architecture.

3.3. Hardware Development Group

• Characteristics: Long iteration cycles, specialized toolchains (EDA, simulation, verification), extremely high error cost, and highly mechanized criteria (formal verification, timing convergence, UVM coverage).
• Key implication: The L5 layer has a natural advantage here — hardware verification already has a mature coverage-driven verification methodology that can be transplanted directly into Harness's evaluation layer; L2's sandbox isolation is especially critical for protecting IP and process libraries.
• What most needs strengthening: L1 context engineering — hardware design context far exceeds code in scale, so compression and indexing strategies decide success or failure.
• Risk: Tool-invocation side effects are irreversible (e.g., tape-out-related operations), so L6 governance must set hard human gates.

3.4. Knowledge Collaboration Group

• Characteristics: The tasks are documents, processes, and collaboration; criteria are highly subjective; failure cost is low but errors spread widely.
• Key implication: The L4 memory and state layer is the core — cross-session knowledge accumulation and retrieval; the L1 layer should invest mainly in Select (precise retrieval) and Isolate (sub-agent isolation).
• What most needs strengthening: L5 evaluation — subjective tasks are exactly the zone where the self-evaluation bias that Anthropic points to is most prevalent, so it must use an independent Evaluator + multi-dimensional weights.
• Risk: Context poisoning does the most harm in knowledge scenarios — once false information enters long-term memory it keeps propagating.

3.5. Data Science Group

• Characteristics: Highly exploratory, requires extensive parallel trial-and-error, some scenarios involve HPC and AI4S, and criteria are partly mechanizable (metrics) and partly not (quality of insights).
• Key implication: The L3 layer should make heavy use of sub-agent dispatch — Anthropic's official parameters (a sub-agent spends tens of thousands of tokens and returns only a 1,000–2,000 token summary) are a natural fit for parallel exploration; the L6 layer needs budget guardrails to keep exploration costs from spiraling out of control.
• What most needs strengthening: The L4 state layer — experiment state, intermediate artifacts, and reproducibility records.
• Risk: Interruption recovery and cost control under long-running execution (analogous to "single tasks exceeding 6 hours").

3.6. Creative Industry Group

• Characteristics: Criteria are highly subjective, there is no binary answer, and outputs are hard to verify automatically.
• Key implication: This is the group among all six that most needs "separating the evaluator from the evaluatee" — Anthropic explicitly notes that self-evaluation bias is especially pronounced in "subjective tasks without a binary criterion (e.g., design)." The recommendation is a Planner / Generator / Evaluator three-role setup plus a hard threshold on multi-dimensional weights.
• What most needs strengthening: Human-on-the-loop design of the L5 evaluation layer — not having humans fix the artifacts (in the loop), but having humans fix the Harness that produces the artifacts (on the loop).
• Risk: Incorrectly reducing subjective quality to mechanically verifiable proxy metrics, causing metrics to rise while quality falls.

3.7. Risk & Compliance Group

• Characteristics: Criteria come from external regulations and internal policies; compliance is a hard constraint; audit traceability takes priority over efficiency.
• Key implication: The L6 governance layer is the protagonist, not a supporting player. Three actionable recommendations:

1. Write compliance red lines as mechanically enforced rules (analogous to OpenAI's custom linter), rather than as prompt requirements;
2. Adopt a three-level configuration scope (project-level / user-level / organization-level), with organization-level managed settings used for mandatory rollout;
3. Credential protection (file deny / environment-variable deny / mask) should be the default configuration, not an option.

• Most critical to watch: Parts 2 and 3 of GB/Z 185—2026 (identity codes and identity management) — the authorities have made clear they will "advance the conversion of identity-code-related standards into mandatory national standards in due course," and are accelerating the development of agent audit standards. Harness products operating within China should prepare early for identity-code compatibility and audit-trail capabilities.
• Risk: Conflating observability (L5) with auditing (L6) — the former serves optimization, the latter serves evidence; the two differ in retention period, tamper-resistance requirements, and access subjects.

4. Assessment of the Market Landscape

4.1. A Four-Layer Market Structure

4.2. Three Main Competitive Routes

4.3. Competitive Landscape Assessment

Judgment One · Model vendors extending into Harness is the mainstream direction.

Both Anthropic and OpenAI already provide models and Harness together; there are also signs of domestic vendors forming Harness teams (Grade C, ). The technical basis for this route is the post-training coupling between model and Harness — whoever controls the Harness controls the trajectory data, and trajectory data is becoming a competitive moat (Grade B).

Judgment Two · Neutral governance of the protocol layer is irreversible, but there is a risk that "de facto standards outweigh legal standards."

Linux Foundation Executive Director Jim Zemlin has publicly acknowledged this risk, drawing the Kubernetes analogy that "dominance emerges from merit and not vendor control." The practical implication for adopters is: even with open standards, prepare for "one particular implementation becoming the default."

Judgment Three · The Chinese market will see a compliance threshold at the identity level.

GB/Z 185—2026 takes the identity code as its core lever, has already planned conversion to mandatory national standards, and is simultaneously accelerating the drafting of agent-audit and agent-transaction standards. Assessment: within China, "whether one can connect to the national agent identity system" will become one of the admission conditions for Harness products, which is markedly different from the protocol-interoperability-centric competitive dimension in Europe and the US.

Judgment Four · There is no authoritative measurement of the market size of the Harness layer itself.

This research did not find authoritative market-size data for the Harness layer itself, so it is marked [待填写]. Existing generative-AI market forecasts differ by nearly an order of magnitude (Grand View Research about $324.7 billion in 2033 vs. Bloomberg Intelligence about $2.3 trillion in 2032, both ), and are not recommended as a basis for decisions.

5. Open Questions

The following questions are ones that vendors and researchers have publicly admitted are unsolved, yet they precisely define the agenda for the next generation of Harness. This project does not offer speculative answers; it only records them faithfully:

From Anthropic officially (Grade A)

1. Is a single general-purpose coding agent always optimal, or is a multi-agent architecture split into specialized agents for testing, QA, cleanup, etc. better — the authorities frankly admit "it is not yet clear."
2. Can these lessons generalize to other long-horizon tasks such as scientific research or financial modeling — the authorities say "the answer is still being explored."

From martinfowler.com (Grade A)

1. As a Harness grows, how do we keep guides (feed-forward control) and sensors (feedback control) in sync without contradicting each other?
2. When instructions conflict with feedback signals, to what extent can we trust the agent's trade-offs?
3. A sensor that never fires — is that a sign of high quality, or a sign of inadequate detection?
4. We need Harness coverage and quality-evaluation methods analogous to code coverage and mutation testing.

From Dex Horthy (Grade B)

1. The best eval may be "having the model, without knowing what the next step is, continuously build 20 features in a single codebase."

From this project's observation (assessment)

1. Once the model is deeply post-training coupled to a particular Harness structure, how will the cost of "swapping out the Harness" change? This determines whether the future Harness is replaceable middleware or locked-in infrastructure.
2. Will the two identity systems of China and the US (A2A Signed Agent Cards and the GB/Z identity code) recognize each other? This determines whether the global agent network is one network or two.

6. Conclusion

Reviewing the argument chain composed of the five documents in this module:

• 01-Introduction establishes what AI Harness is — the carrying layer that turns the model's uncertainty into engineering predictability, with the six-layer capability model as its anatomy;
• 02-Development History explains where it came from — the bottleneck has kept moving outward from inside the model, and each generation settles the previous generation's manual work into default capability;
• 03-Architecture Evolution describes what it is like now — four pillars of context engineering, sandboxed execution, trajectory evaluation, and permission governance, with evidence already at the million-line scale;
• 04-Future Development assesses where it is going — technically becoming lighter but not disappearing, the standardization window closing, and the trust gap in industry being the market opportunity;
• 05-Summary converges into assessments and implications — eight core propositions, differentiated paths for the six industry groups, and a landscape judgment of the four-layer market structure.

If only three sentences could remain, they would be:

1. The bottleneck has already moved from inside the model to outside the model — this is not rhetoric, but the same judgment reached by two leading labs in early 2026.
2. Constraint is capability — the 84% reduction in permission prompts and the 3.5 PRs/person/day throughput both come from better constraints, not fewer ones.
3. We are in the window where standardization is closing while the methodology has not yet converged — the stretch most worth seizing between the past three years and the next three.

7. Information-Gap Statement

This document is a summary; all of its arguments are drawn from the preceding four documents, so it fully inherits their information gaps. Here we only aggregate the seven most critical ones:

1. International standards for agent interconnection at the ISO/IEC level: no published or registered standard number was found; there is no authoritative information. This directly affects the completeness of the global market-landscape judgment in Section 4.3.
2. The originator and first occurrence of the term "Agent Harness": no definitive primary source was found; there is no authoritative information.
3. The original text and publication date of LangChain's "The Anatomy of an Agent Harness": no original text was found; there is no authoritative information.
4. Official Terminal-Bench leaderboard data: the official site was not scraped directly; this document does not cite specific leaderboard figures — the relevant numbers are in 03-架构演进.md and marked .
5. Official URLs for DORA 2025 and McKinsey's "The State of AI 2025": accessibility was not verified; the DORA conclusion cited in Section 2.6 is marked .
6. The to-be-verified figures cited in this document: Codex Auto-review's 1/200 and 99%; the Vercel tool-trimming experiment; domestic vendor moves (DeepSeek, Xiaomi, Lingxi Zhiyong); and the 2026 release timeline of the various Claude versions.
7. The market size of the Harness layer itself: this research found no authoritative measurement, so it is marked [待填写].

8. References

1. Harness design for long-running application development — Anthropic, 2026. https://www.anthropic.com/engineering/harness-design-long-running-apps
2. Effective harnesses for long-running agents — Anthropic, 2025. https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
3. Effective context engineering for AI agents — Anthropic, 2025. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
4. Sandboxing: a safer and more autonomous approach — Anthropic, 2025. https://www.anthropic.com/engineering/claude-code-sandboxing
5. Harness engineering: leveraging Codex in an agent-first world — OpenAI, 2026-02-11. https://openai.com/index/harness-engineering/
6. Harness engineering for coding agent users — Birgitta Böckeler, martinfowler.com, 2026. https://martinfowler.com/articles/harness-engineering.html
7. Humans and Agents in Software Engineering Loops — martinfowler.com, 2026. https://martinfowler.com/articles/exploring-gen-ai/humans-and-agents.html
8. My AI Adoption Journey — Mitchell Hashimoto, 2026-02-05. https://mitchellh.com/writing/my-ai-adoption-journey
9. Linux Foundation Announces the Formation of the AAIF — Linux Foundation / AAIF, 2025-12-09. https://aaif.io/
10. A year of open collaboration: Celebrating the anniversary of A2A — Google Open Source Blog, 2026-04-16. https://opensource.googleblog.com/
11. 2025 Stack Overflow Developer Survey — Stack Overflow, 2025-07-29. https://survey.stackoverflow.co/2025/
12. AI Adoption Stats & Trends (aggregating DORA / McKinsey / JetBrains) — daily.dev, 2026-07-19. https://daily.dev/agentic-ai-hub/ai-adoption-stats-trends/
13. SWE-bench official site — Princeton / community. https://www.swebench.com/
14. Terminal-Bench — Stanford / Laude Institute. https://www.tbench.ai/
15. "Artificial Intelligence Agent Interconnection" National Standard Series (Interpretation) — China Industrial Economic Information Network, 2026. https://cinic.org.cn/xw/zcdt/1643418.html
16. "Artificial Intelligence Agent Interconnection" National Standard Series (Release report) — People's Daily Online, 2026-07-09. https://finance-app.people.cn/n1/2026/0709/c1004-40757059.html
17. GB/Z 185—2026 roll-out: first batch of Yangtze River Delta agent identity-code nodes issued — China Daily, 2026-09-04. https://cn.chinadaily.com.cn/a/202609/04/WS6a9a6773e4b09a165c788098.html
18. Agent Harness: the core paradigm of AI engineering in 2026 — Tencent Cloud Developer Community, 2026. https://developer.cloud.tencent.com/article/2698416
19. Model Context Protocol official site and specification — MCP / AAIF. https://modelcontextprotocol.io/
20. MCP Protocol Versions — MCP Ruby SDK, 2026. https://ruby.sdk.modelcontextprotocol.io/protocol-versions/